Prediction model generating device, prediction device, prediction model generating method, prediction method, and program

Abstract

To predict whether or not a pain occurs in a prediction object person consecutively in term of time.SOLUTION: A prediction model generating device generates a prediction model of outputting a value as a prediction reference about whether or not a pain occurs in a prediction object person, by inputting data relating to an electrocardiogram of the prediction object person, with data relating to an electrocardiogram of a patient when no pain occurs in the patient as teacher data.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a prediction model generating device, a prediction device, a prediction model generating method, a prediction method, and a program. [Background technology]

[0002] Many studies have been conducted that attempt to objectively evaluate pain, which is a subjective sensation (for example, Non-Patent Documents 1 and 2). In studies that attempt to objectively evaluate pain, pain is mainly evaluated from physiological changes that occur when heat or electrical stimulation is generated. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] E, K, Naeini et al., “Pain Recognition With Electrocardiographic Features in Postoperative Patients: Method Validation Study” J Med Internet Res, vol. 23, no.5, pp. 1-13, 2021. [Non-Patent Document 2] SAH Aqajari et al., “Pain Assessment Tool With Electrodermal Activity for Postoperative Patients: Method Validation Study”, JMIR Mhealth Uhealth, vol.9, no. 5, pp. 1-11, 2021 Summary of the Invention [Problem to be solved by the invention]

[0004] However, conventional research has not been able to predict in real time whether pain will occur. Non-Patent Documents 1 and 2 also rely on self-assessment scales to evaluate pain, and only allow discontinuous evaluation over time.

[0005] An object of the present invention is to provide a prediction model generation device, a prediction device, a prediction model generation method, a prediction method, and a program that can predict whether a subject will experience pain continuously over time. [Means for solving the problem]

[0006] One aspect of the present invention is a predictive model generation device that uses data regarding a patient's electrocardiogram when the patient is not experiencing pain as teacher data, and data regarding a subject's electrocardiogram as input, and generates a predictive model that outputs a value that serves as a prediction standard for whether or not the subject will experience pain.

[0007] One aspect of the present invention is a prediction device that predicts whether a subject will experience pain by inputting data regarding the patient's electrocardiogram into a prediction model that is trained to input data regarding the patient's electrocardiogram and output a value that serves as a prediction standard for whether the patient will experience pain. Effect of the Invention

[0008] According to the present invention, it is possible to predict whether or not a subject will experience pain continuously over time. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 is a diagram illustrating an example of a configuration of a prediction model generating device according to an embodiment of the present invention. [Diagram 2] FIG. 1 is a diagram illustrating an example of a configuration of a teacher data generation system. [Diagram 3] FIG. 13 is a diagram showing electrocardiograms to be excluded. [Figure 4] 4 is a flowchart showing the operation of the predictive model generating device of the present embodiment. [Diagram 5] FIG. 1 is a diagram illustrating an example of a configuration of a prediction system according to an embodiment of the present invention. [Figure 6] FIG. 1 illustrates an example of a configuration of a prediction device according to an embodiment of the present invention. [Figure 7] 10 is a flowchart illustrating an example of an operation of the prediction device according to the present embodiment. [Figure 8] FIG. 13 is a diagram showing the results of pain prediction made by a prediction device for a certain prediction subject, and the time point at which the pain occurred. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] (Prediction model generation device) Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. 1 is a diagram illustrating an example of the configuration of a prediction model generation device 1 according to the present embodiment. The prediction model generation device 1 generates a prediction model. The prediction model is a model that predicts whether or not a person who is a target of prediction will experience pain. The prediction model generating device 1 includes a teacher data acquiring unit 11, a prediction model generating unit 12, and a prediction model output unit 13.

[0011] The teacher data acquiring unit 11 acquires teacher data. The teacher data is data related to the electrocardiogram of the patient when the patient is not in pain. 2 is a diagram showing an example of the configuration of the teacher data generation system 2. The teacher data generation system 2 generates teacher data of a patient P which is acquired by a teacher data acquisition unit 11. The teacher data generation system 2 includes an electrocardiogram measuring device 21, a signal output device 22, and a teacher data generation device 23.

[0012] The electrocardiogram measuring device 21 measures the electrocardiogram of the patient P. The electrocardiogram measuring device 21 outputs the electrocardiogram measured in the patient P to the teacher data generating device 23. The electrocardiogram measuring device 21 may immediately output the measured electrocardiogram to the teacher data generating device 23, or may temporarily store the measured electrocardiogram and periodically output it to the teacher data generating device 23.

[0013] The signal output device 22 outputs a signal indicating the presence of pain to the teacher data generating device 23. The signal output device 22 includes, for example, a button, and outputs a signal to the teacher data generating device 23 when the button is pressed. The patient P uses the signal output device 22 and presses the button while feeling pain, thereby outputting a signal indicating the presence of pain to the teacher data generating device 23.

[0014] The signal output device 22 is, for example, a device used in intravenous patient controlled analgesia (IV-PCA) that administers a painkiller when a button is pressed (hereinafter, referred to as a PCA pump). The signal output device 22 may be an electronic device such as a smartphone, and may output a signal indicating the presence of pain to the teacher data generation device 23 when the screen is tapped.

[0015] The teacher data generating device 23 generates teacher data based on the patient P's electrocardiogram and a signal indicating the presence of pain. The teacher data generating device 23 synchronizes the electrocardiogram of the patient P with the signal indicating the pain. The teacher data generating device 23 excludes electrocardiograms within a predetermined time before and after the time when the signal indicating the pain is received from the electrocardiogram of the patient P, and generates an electrocardiogram when the patient P is not experiencing pain or when pain does not occur in a similar time period (hereinafter referred to as a pain-free electrocardiogram). Fig. 3 is a diagram showing the electrocardiograms to be excluded. The teacher data generating device 23 may generate a pain-free electrocardiogram by another method, such as excluding an electrocardiogram while a signal indicating pain is being received.

[0016] The teacher data generating device 23 generates data related to the electrocardiogram based on the painless electrocardiogram of the patient P. The data related to the electrocardiogram is data indicating an index (hereinafter referred to as HRV index) characterizing the heart rate variability (HRV) calculated from the interval between R waves (RR Interval, RRI) in an electrocardiogram. The HRV index includes, for example, SDNN which is the standard deviation of the RRI, RMSSD which is the root mean square of the difference between adjacent RRIs, Total Power which is the variance of the RRI, LF which is the value obtained by dividing the power in the low frequency region (e.g., 0.04-0.15 Hz) by Total Power, HF which is the value obtained by dividing the power in the high frequency region (e.g., 0.15-0.4 Hz) by Total Power, or LF / HF which is the ratio of LF to HF.

[0017] The electrocardiogram data may be data indicating the RRI. The electrocardiogram data may be image data of an electrocardiogram. In this manner, the teacher data generation system 2 generates data relating to pain-free electrocardiograms as teacher data.

[0018] The teacher data generating device 23 may generate data relating to the electrocardiogram based on the electrocardiogram of the patient P, including the electrocardiogram when the patient P is in pain, in association with the timing of receiving a signal indicating the presence of pain. The data is used as threshold adjustment data for setting a prediction threshold of a prediction model, which will be described later.

[0019] Returning now to FIG. 1, the prediction model generating device 1 will be described. The prediction model generating unit 12 generates a prediction model by machine learning using teacher data. The prediction model is a model that receives data on the electrocardiogram of a prediction subject as an input and outputs a value that serves as a prediction criterion for whether or not the prediction subject experiences pain.

[0020] The prediction model is a model that takes into account the time series of electrocardiograms. For example, the prediction model is an autoencoder model with self-attention (SA-AE) that incorporates a self-attention mechanism (SA) in the hidden layer of an autoencoder (AE).

[0021] The prediction model generation unit 12 generates a prediction model by inputting, for example, training data generated from a patient P different from the training data used to generate the prediction model, calculating the mean square error (MSE) between the input data and the output data, and updating the parameters of the prediction model so as to minimize the mean square error. In this way, a prediction model is generated in which the RMSE between the input data and the data output when data related to the electrocardiogram of the patient P in no pain is input is minimized.

[0022] The prediction model calculates, for example, the root mean square error (RMSE) between input data relating to the electrocardiogram of the prediction subject and teacher data relating to the electrocardiogram of the patient P when the patient P is not in pain as a reconstruction error (RE). Based on the calculated RE, a prediction device 32, which will be described later, predicts whether or not the prediction subject will experience pain. The prediction model may predict whether the predicted subject will experience pain based on the calculated RE. For example, the prediction model may predict that the predicted subject will experience pain when the RE is equal to or greater than a predetermined threshold for a predetermined period of time.

[0023] The prediction model output unit 13 outputs the generated prediction model. The output prediction model is stored in, for example, a storage unit 329 of the prediction device 32, which will be described later.

[0024] 4 is a flowchart showing the operation of the prediction model generating device 1 of this embodiment. The teacher data acquiring unit 11 acquires teacher data (step S11). The prediction model generating unit 12 generates a prediction model by machine learning using the teacher data (step S12). The prediction model output unit 13 outputs the generated prediction model (step S13).

[0025] (Prediction System) 5 is a diagram showing an example of the configuration of the prediction system 3 of this embodiment. The prediction system 3 predicts whether or not the prediction target S has pain, based on data related to the electrocardiogram of the prediction target S. The prediction system 3 includes an electrocardiogram measuring device 31 and a prediction device 32. The electrocardiogram measuring device 31 is a device having the same functions as the electrocardiogram measuring device 21 in the teacher data generation system 2. The electrocardiogram measuring device 31 measures the electrocardiogram of the prediction subject S and outputs it to the prediction device 32. The prediction device 32 predicts whether or not the prediction subject S is in pain based on the electrocardiogram of the prediction subject S. 6 is a diagram showing an example of the configuration of the prediction device 32 of this embodiment. The prediction device 32 includes an electrocardiogram acquiring unit 321, an electrocardiogram conversion unit 322, a prediction unit 323, a prediction result output unit 324, and a storage unit 329. The storage unit 329 stores the prediction model output by the prediction model generation device 1.

[0026] The electrocardiogram acquiring unit 321 acquires an electrocardiogram of the prediction target S. The electrocardiogram acquiring unit 321 records the acquired electrocardiogram in the storage unit 329. The electrocardiogram stored in the storage unit 329 may be an electrocardiogram at a most recent predetermined time.

[0027] The electrocardiogram conversion unit 322 converts the electrocardiogram of the prediction target S to generate data related to the electrocardiogram. The electrocardiogram data generated by the electrocardiogram conversion unit 322 is the same type of data as the electrocardiogram data of the teacher data used to generate the prediction model. For example, when the electrocardiogram data of the teacher data is data indicating an HRV index, the electrocardiogram conversion unit 322 calculates the HRV index based on the electrocardiogram of the prediction target S to generate data indicating the HRV index.

[0028] The electrocardiogram conversion unit 322 converts the electrocardiogram of the prediction target S at the most recent predetermined time to generate data related to the electrocardiogram. In this way, the electrocardiogram conversion unit 322 generates data related to the electrocardiogram of the prediction target S at each time point.

[0029] The prediction unit 323 calculates the RE by inputting the data related to the electrocardiogram into a prediction model stored in the storage unit 329. The prediction unit 323 calculates the RE every time the data related to the electrocardiogram is updated. The prediction unit 323 predicts that the prediction subject will experience pain when the calculated RE is equal to or greater than the RE threshold for a predetermined time. Here, the predetermined time is set by the prediction device 32 to an arbitrary time.

[0030] The RE threshold is adjusted, for example, depending on the prediction accuracy of the prediction model. For example, electrocardiogram data included in the threshold adjustment data of a patient in the prediction model is input, and it is predicted that the patient will experience pain when the calculated RE is equal to or greater than the set RE threshold for a predetermined period of time. The threshold adjustment data includes data indicating the timing of receiving a signal indicating the presence of pain, i.e., data indicating the timing when the patient experienced pain, so the prediction result is compared with the data indicating the timing when the pain occurred to calculate the prediction accuracy. The RE threshold is adjusted by changing the RE threshold and calculating the prediction accuracy.

[0031] Prediction accuracy is the difference between sensitivity (TPR) and false positive rate (FPR). The RE threshold is adjusted to maximize the difference between TPR and FPR. TPR is the proportion of times that the prediction model predicts that the predicted subject will have pain a certain time before the time when the signal indicating pain is received. FPR is the number of times that the prediction model predicts that the predicted subject will have pain at a time other than the certain time before or after the time when the signal indicating pain is received. The RE threshold is set in the above manner. Note that the RE threshold may be set to a value smaller than the RE threshold when the prediction accuracy is maximized (for example, a value that is 90-95% of the RE threshold).

[0032] The predetermined time and the RE threshold may be set based on the characteristics of the predicted subject S. If the predicted subject S uses a PCA pump, the predetermined time and the RE threshold may be set based on the frequency of use, such as the time when the PCA pump is not pressed.

[0033] The prediction result output unit 324 outputs the prediction result by the prediction unit 323. The prediction result is output to, for example, an external display device and displayed. The prediction result is output to, for example, an alert device, and if the prediction result indicates that pain will occur, the alert device issues an alert.

[0034] 7 is a flowchart showing an example of the operation of the prediction device 32 of this embodiment. The electrocardiogram acquisition unit 321 acquires the electrocardiogram of the prediction target S (step S31). The electrocardiogram conversion unit 322 converts the electrocardiogram of the prediction target S to generate data related to the electrocardiogram (step S32). The prediction unit 323 inputs the data related to the electrocardiogram into a prediction model to predict whether the prediction target S is in pain (step S33). The prediction result output unit 324 outputs the prediction result (step S34).

[0035] As a result, the prediction device 32 can predict whether or not the prediction subject S will experience pain. The prediction device 32 calculates data related to the electrocardiogram based on the electrocardiogram of the prediction subject measured in real time by the electrocardiogram measuring device 31, inputs the data related to the electrocardiogram into a prediction model, and predicts whether or not the prediction subject will experience pain based on the output value. This allows the prediction device 32 to continuously predict whether or not the prediction subject will experience pain over time.

[0036] (Measurement results) The following describes the verification performed on the prediction model generating device 1 and the prediction device 32. First, the teacher data generating system 2 generated teacher data and threshold adjustment data. In generating teacher data, the teacher data generating device 23 calculated data related to the electrocardiogram from the electrocardiogram excluding the electrocardiogram 60 minutes before and after the time when the signal was acquired from the signal output device 22. A PCA pump was used as the signal output device 22. As a result, the teacher data generating device 23 generated data related to the electrocardiogram based on the electrocardiogram when the patient P was not in pain. In addition, the teacher data generating device 23 calculated six indices, SDNN, RMSSD, Total Power, LF, HF, and LF / HF, as data related to the electrocardiogram. The teacher data generating device 23 calculated data related to the electrocardiogram from the electrocardiogram that was not removed, and generated the data related to the electrocardiogram and the data at the time when the signal was acquired from the signal output device 22 as threshold adjustment data. The teacher data generation system 2 generated teacher data and threshold adjustment data from eight patients. The patients were divided into seven and one, and a prediction model was generated and verified. The prediction model was generated using training data generated from the seven patients, and the RE threshold, which is the prediction standard for the prediction model, was set using threshold adjustment data generated from the seven patients. The prediction model was verified using threshold adjustment data generated from the remaining one patient.

[0037] The prediction model used was SA-AE. The prediction model was trained by training the model so that the MSE of the input / output error of SA-AE was minimized. To explain using a formula, the model was trained so that L in equation (1) was minimized.

number

[0038] In formula (1), k is the number of features, x k、i is the input feature, n is the length of the input sequence, and x^ k、i is the feature output from the prediction model. Six HRV indices with a sequence length of 60 were used as input features. The model's hyperparameters were 5 heads, 6 SA layers, and a dropout rate of 20%, and the model was trained using the Adam optimizer with a learning rate of 0.002. The prediction model calculated RMSE as RE. RE was calculated using Equation (2).

number

[0039] The RE threshold, which is the prediction standard of the prediction model, was determined using a grid search so as to maximize the difference between the TPR and the FPR. The TPR was the percentage of times that the prediction device 32 predicted that pain would occur from the time point when the signal indicating pain was received until 15 minutes before. The FPR was the number of times per hour that the prediction model predicted that pain would occur outside of the predetermined time before and after the time point when the signal indicating pain was received.

[0040] In the prediction device 32, the electrocardiogram conversion unit 322 calculated six indices, SDNN, RMSSD, Total Power, LF, HF, and LF / HF, based on the RRI for three minutes. The prediction unit 323 then inputted the six indices into a prediction model to calculate RE. If the calculated RE exceeds 90-95% of the determined RE threshold continuously for 5-15 seconds, it is predicted that the prediction subject will experience pain. Fig. 8 is a diagram showing the results of pain prediction by prediction device 32 for a certain prediction subject, and the time point at which the pain occurred. The graph shown in Fig. 8 has RE on the vertical axis and time on the horizontal axis. The prediction device was able to predict that the pain would occur approximately 10 minutes before the time point at which the pain occurred (the time point at which the PCA pump was pressed).

[0041] Other Embodiments Although one embodiment of the present invention has been described in detail above with reference to the drawings, the specific configuration is not limited to the above, and various design changes, etc. are possible within the scope that does not deviate from the gist of the present invention.

[0042] The prediction model generating device 1, the teacher data generating device 23, and the prediction device 32 in the above-mentioned embodiment may be partly or entirely realized by a computer. In that case, a program for realizing this function may be recorded in a computer-readable recording medium, and the program recorded in the recording medium may be read into a computer system and executed to realize the function. The term "computer system" as used herein includes the OS and hardware of peripheral devices. The term "computer-readable recording medium" refers to portable media such as flexible disks, optical magnetic disks, ROMs, and CD-ROMs, and recording devices such as hard disks built into a computer system. The term "computer-readable recording medium" may also include a medium that dynamically holds a program for a short period of time, such as a communication line when a program is transmitted via a network such as the Internet or a communication line such as a telephone line, and a medium that holds a program for a certain period of time, such as a volatile memory inside a computer system that is a server or client in that case. The above-mentioned program may be a program for realizing part of the above-mentioned function, or may be a program that can realize the above-mentioned function in combination with a program already recorded in the computer system. Furthermore, the prediction model generating device 1, the teacher data generating device 23, and the prediction device 32 may be partly or entirely realized using a programmable logic device such as an FPGA (Field Programmable Gate Array). [Explanation of symbols]

[0043] 1 Prediction model generating device, 11 Teacher data acquisition unit, 12 Prediction model generating unit, 13 Prediction model output unit, 2 Teacher data generating system, 21 Electrocardiogram measuring device, 22 Signal output device, 23 Teacher data generating device, 3 Prediction system, 31 Electrocardiogram measuring device, 32 Prediction device, 321 Electrocardiogram acquiring unit, 322 Electrocardiogram conversion unit, 323 Prediction unit, 324 Prediction result output unit, 329 Memory unit

Claims

1. Data on the electrocardiogram of the patient when the patient is not experiencing pain is used as training data, A prediction model is generated that receives data on the electrocardiogram of a prediction subject and outputs a value that serves as a prediction criterion for whether or not the prediction subject will experience pain. Predictive model generator.

2. The prediction model calculates a reconstruction error between the data on the electrocardiogram of the prediction subject and the data on the electrocardiogram of the patient when the patient is not in pain; When the reconstruction error is equal to or greater than a predetermined value for a predetermined time, it is predicted that the subject will experience pain. The predictive model generating device according to claim 1 .

3. The electrocardiogram data is an index characterizing heart rate variability calculated from an interval between R waves in the electrocardiogram. The prediction model generating device according to claim 1 .

4. A prediction model is trained to receive data on the patient's electrocardiogram as an input and output a value that serves as a prediction criterion for whether the patient will experience pain, and the prediction model predicts whether the patient will experience pain by inputting data on the electrocardiogram of the patient. Prediction device.

5. By inputting data on the electrocardiogram of the prediction subject into the prediction model, a reconstruction error is calculated between the data on the electrocardiogram of the prediction subject and data on the electrocardiogram of the patient when the patient is not in pain; When the reconstruction error is equal to or greater than a predetermined value for a predetermined time, it is predicted that the subject will experience pain. The prediction device according to claim 4 .

6. Data on the electrocardiogram of the patient when the patient is not experiencing pain is used as training data, A prediction model is generated that receives data on the electrocardiogram of a prediction subject and outputs a value that serves as a prediction criterion for whether or not the prediction subject will experience pain. Methods for generating predictive models.

7. A prediction model is trained to receive data on the patient's electrocardiogram as an input and output a value that serves as a prediction criterion for whether the patient will experience pain, and the prediction model predicts whether the patient will experience pain by inputting data on the electrocardiogram of the patient. Forecasting methods.

8. A method for generating a predictive model according to claim 6, program.

9. A method for predicting a prediction result according to claim 7, program.

Images